Aegis

The programmable multi-agent meta-harness.

Drives Claude Code, Gemini CLI, and OpenCode in one terminal, gives them six primitives for working together, and lets you orchestrate them with deterministic Python workflows and scheduled jobs.

┌ aegis · 3 agents · ~/code/aegis ─────────────────────────────────────┐
│ ● 1 lucid-knuth ·opus·   ● 2 wry-hopper ·gemini·   ● 3 brisk-curie * │
│                                                                       │
│ › explain the retry path in worker.py                                 │
│                                                                       │
│ ⠹ Thinking… (3.2s)                                                    │
│ ⏺ Read(worker.py)                                                     │
│   └ ok                                                                 │
│ The retry path lives in _run_turn at line 142 …                       │
│                                                                       │
│ ⏺ aegis_handoff(target=wry-hopper)                                    │
│   └ delivered to wry-hopper                                           │
│                                                                       │
│ queues: tests ●1/2 ○0 ✓3 ✗0    last: brisk-curie                      │
│ lucid-knuth ·opus· opus·full   ↑128k (94% cached) ↓1k                 │
│ ───────────────────────────────────────────────────────────────────── │
│ › ask something…                                                      │
└───────────────────────────────────────────────────────────────────────┘

What aegis is

Meta-harness. Most agentic frameworks (CrewAI, LangGraph, AutoGen, the long list) talk directly to LLM providers — they replace your coding agent and reimplement tool use, permissions, sandboxing, terminal integration. Aegis sits above your existing coding agents and drives them over their structured protocols — stream-json for Claude Code, the Agent Client Protocol (ACP) for Gemini CLI and OpenCode, with a clean driver seam for whatever lands next. The harness keeps owning tool use, model selection, MCP hosting, sandboxing. Aegis owns the layer above — tabs, routing, delegation, persistence — the things a single-conversation CLI was never built to do.

Multi-agent. Six composable coordination primitives, all wired into one MCP plane every spawned agent sees. Inbox for fire-and-forget context handoff. Queue for spawn-a-worker-on- demand dispatch. Canvas for shared markdown blackboards. Terminal for live shared PTYs. Groups for broadcast-and- gather across a committee. Workflow for deterministic Python orchestration. Mix providers freely — a Claude tab hands off to a Gemini tab; an OpenCode worker drops its result in your inbox; three agents co-author a canvas; a workflow drives all of them in lockstep.

Programmable. The substrate is scriptable from the outside and the inside. Outside: @workflow-decorated Python functions that compose the six primitives into deterministic procedures (TDD loops, branch reviews, spec-to-plan-to-implementation pipelines), cron- scheduled or invoked on demand. Inside: spawned agents can mutate .aegis.yaml themselves through MCP — declare a new specialised agent profile, register a queue, drop a plugin dir, and dispatch work to it in the same session, no restart. The substrate extends from inside.

Six primitives for agent coordination

— the multi-agent pillar.

Each primitive has one verb and lands the same way in the receiving agent's transcript: as a ✉ block with a sender tag, timestamp, and a short body preview. One delivery channel, six wake patterns.

→ Inbox — send context to a peer

Any agent can hand off to any other live agent. Fire-and-forget; the recipient gets a normal user-message turn tagged with the sender's handle. Use when you want a specific peer to pick up where you left off.

aegis_handoff(target_handle="reviewer", from_handle="impl",
              context="PR ready at branch feat/x — please review")
# → reviewer's transcript:
#   ✉ from agent:impl · 17:42:03Z
#     PR ready at branch feat/x — please review

⏳ Queue — spawn a worker on demand

Enqueue a task to a named queue and the substrate spawns a fresh agent of the queue's configured profile, runs the payload as its opening turn, and (with callback=true) delivers the worker's final result back to your inbox. Producer keeps working between enqueue and callback. Generalizes delegation: parallelism, max-in-flight caps, restart safety, all built in.

aegis_enqueue(queue="review", payload="…full self-contained prompt…",
              from_handle="impl", callback=True)
# → {task_id: 01HK…, queued_position: 1}
# …minutes later, in impl's transcript:
#   ✉ from queue:review · task#01HK… · ok · 17:46:11Z
#     PR looks clean. Two nits flagged in the diff comments…

▦ Canvas — collaborate on a shared document

Open a shared markdown file. Multiple agents read it, write sections of it, subscribe to it. Each write wakes every other subscriber with a diff-aware notification. The classical blackboard pattern — terminal- native, MCP-driven, file-backed (you can grep it, commit it, open it in your editor).

# PM
aegis_canvas_open(name="report-q3", file="vault/reports/q3.md",
                  from_handle="pm")
aegis_canvas_subscribe(name="report-q3", from_handle="pm")

# Researcher (in another tab, after a handoff)
aegis_canvas_write_section(name="report-q3", section="data",
                           content="Q3 numbers came in stronger…",
                           from_handle="researcher")
# → PM's transcript:
#   ✉ from canvas:report-q3 · 20:30:00Z
#     section "data" · written by agent:researcher (+18 / -3 lines)
#     ──
#     Q3 numbers came in stronger than projected…

▮ Terminal — share a live shell

Spawn a PTY-backed shell that any agent (or Alex) can run commands on, send raw keystrokes to, and subscribe to. Command boundaries are detected from OSC 133 shell-integration markers; every finalized command lands in an append-only JSONL ledger and wakes subscribers through the same ✉ channel.

# PM
aegis_term_spawn(name="build", from_handle="pm")
aegis_term_subscribe(name="build", from_handle="pm")

# builder (after a handoff)
rec = aegis_term_run(name="build", cmd="pytest -q",
                     from_handle="builder")
# → PM's transcript:
#   ✉ from term:build · 14:03:25Z
#     $ pytest -q  · run by agent:builder
#     exit 0 · 4.20s
#     ──
#     6 passed in 4.18s

▣ Groups — broadcast and gather

Form a named committee of agents that share one inbox-fanout channel and one in-flight broadcast slot. Send one structured four-field question — objective, output_format, tool_guidance, boundaries — collect N parallel answers, reduce them into a single result. Use when the same question has multiple useful perspectives, or when you want to race providers and keep the fastest.

# spawn three reviewers with different lenses
aegis_group_spawn_mixed(name="audit", from_handle="pm",
    profiles=["sec_reviewer", "style_reviewer", "logic_reviewer"])

aegis_group_broadcast(name="audit",
    objective="audit PR #214 (branch feat/rate-limit)",
    output_format="bullet list, each item severity-tagged (high/med/low)",
    tool_guidance="prefer Read + Grep; avoid Bash and Edit",
    boundaries="report only — no patches, no commits")

# collect every reply, keyed by reviewer
result = aegis_group_wait_all(name="audit",
                              timeout=300,
                              reducer="join_by_handle")
# → result.reduced = {"sec_reviewer": "…", "style_reviewer": "…",
#                     "logic_reviewer": "…"}

Switching to aegis_group_wait_any returns on the first reply and (by default) sends a passive cancel envelope to the losers — useful when the cheapest acceptable answer wins. Built-in reducers: concat, join_by_handle, last_wins, majority_vote; custom reducers register one function. Groups also have YAML presets in .aegis.yaml (groups.presets.<name>.profiles: […]) and a dedicated TUI tab with Members / Current broadcast / Recent broadcasts panels.

Reach for it when: multi-lens code audit, fastest-answer racing, cross-provider consensus, generate-and-pick (N candidates → one), role-persona panels (PM / eng / UX react to the same proposal). Full walk-through in docs/groups.md.

⟳ Workflow — deterministic Python orchestration

When the dance has to be reliable — TDD loops, bug triage, multi-step plans, anything where retries with feedback matter — wrap it in a workflow. Plain Python at the top of the stack. Calls agents, runs bash predicates, retries with feedback, captures structured output.

@workflow("tdd-cycle")
async def tdd_cycle(engine, *, feature: str) -> str:
    impl = await engine.spawn("implementer")
    await engine.send(impl, f"Write a failing test for: {feature}")
    await engine.bash_predicate(
        f"pytest tests/ -k {feature} 2>&1 | grep -E 'FAIL|ERROR'",
        retry_with="The test should fail because the feature isn't built yet")
    await engine.send(impl, "Now implement it.")
    await engine.bash_predicate(
        f"pytest tests/ -k {feature}",
        retry_with="Tests are still failing. Output:\n{stdout}")
    reviewer = await engine.spawn("reviewer")
    return await engine.send(reviewer, "Final review of branch.")

Triggered by any agent: aegis_run_workflow(name="tdd-cycle", kwargs={"feature": "rate_limit"}). Workflows sit at the top of the stack — they span agents, they own the loop, they're the right tool when the spec is "follow this exact procedure" rather than "figure it out."

The aegis.workflows package ships four seed workflows registered on import: brainstorm_to_spec (Q/A → spec doc), execute_plan (parse plan → dispatch implementer per task with durable resume), review_branch (parallel reviewer fan-out → report), and tdd_cycle (predicate-driven TDD loop). See docs/workflows.md.

What else is in the box

• Multi-tab TUI. Generated alliterating handles (lucid-knuth, wry-hopper) for agents, purpose names (build, db) for terminals. State dots, sticky *, terminal bell when a backgrounded agent finishes. Click any block to copy it.
• Honest metrics. True input (incl. cache) with cached %, output, tool calls, per-turn and per-session wall-clock. Provisional while streaming, exact at turn end. Live ctx Nk (P%) segment shows the current turn's true input against the model's context window — Opus 4.x at 1M, Sonnet/Haiku at 200k, Gemini at 1M. No log scraping anywhere.
• Queue dashboard. Always-on one-line strip above the status bar shows live per-queue depth and the most recent in-flight worker. Ctrl+D expands into a full-screen modal with QUEUES / IN-FLIGHT / QUEUED / RECENT bands and a live assistant-text tail.
• File viewer + picker. Ctrl+O opens a fuzzy file picker (typeahead, keyboard nav, top-match preselected) over a background watchdog index; pick a file and it lands in a FileTab — syntax-highlighted read-only view by default, e toggles edit mode, Ctrl+S saves, Escape with unsaved edits prompts to discard. Agents can drop you into the same view via the aegis_view_file MCP tool, and Ctrl+click on a backtick-wrapped filename in any agent response opens it directly.
• Config panel. F2 opens the live .aegis.yaml editor inside the TUI — see agents/queues at a glance, add an agent through a validated modal. Same edit helpers back the scriptable aegis config CLI verbs and a parallel MCP surface (aegis_config_add_agent, …_add_queue, …_add_plugin_dir, …_set_schedule_enabled, plus removes and reads), so agents can extend the substrate from inside — declare a queue and enqueue to it within one session, no restart. Panel, CLI, and MCP all route through the same comment-preserving atomic-write path.
• Session persistence. aegis reopens the last workspace by default — agent tabs, terminal tabs, profiles, order, with each underlying session genuinely resumed (model memory intact). aegis --clean opts out.
• Workflow catalog. aegis.workflows ships four ready-to-use seeds (brainstorm_to_spec, execute_plan, review_branch, tdd_cycle); importing them registers. Engine offers ask_human, explicit checkpoint + durable resume, bash_predicate retry loops, and parallel fan-out.
• Headless + web. aegis serve runs the SessionManager + MCP plane without a TUI; add a web: token to drive the team from an installable, mobile-first web/PWA client (aegis web).
• MCP plane. Every spawned agent is injected with the aegis MCP server: orientation (aegis_meta), session listing, handoff, queue dispatch, canvas ops, terminal ops, group broadcast/gather, workflow invocation. One consistent surface across providers. With --strict-mcp-config, aegis is the only MCP server the spawned agent sees.
• Driver parity. The canonical event surface unifies what every substrate publishes: semantic tool kinds (📖 read / ✏️ edit / ⌬ execute / 🔎 search / ✻ think / 🌐 fetch) with locations + raw inputs, file diffs for edits, plan blocks for TodoWrite / AgentPlanUpdate, mid-turn cost / mode / title telemetry, and end-of-turn stop_reason / cost_usd / per-model attribution. Same render code paths for Claude, Gemini, and OpenCode; opencode's per-token thought stream coalesces by message_id into one block per assistant message.

Plugins — extend aegis without forking

Plugins are how third-party code adds behavior to a session. Three composable primitive shapes, auto-imported from disk, installable from anywhere over gh: registry URLs:

• @hook(event) — fires on harness lifecycle events. pre_turn is the mutator (prepend system text, rewrite the user message, block the turn); post_turn / session_start / session_end are observers. Composes deterministically; timeout-wrapped per call; JSONL-logged.
• @tool — first-class FastMCP tools the spawned agent can call. Schema is auto-generated from type hints + docstring. Reserved names (every built-in aegis_*) are guarded at registration.
• @workflow — orchestrated procedures driven by the WorkflowEngine (delegate, spawn, send/drain, bash, groups). CLI-, MCP-, or scheduler-invoked.

A plugin is one directory: a plugin.toml manifest, the Python module(s) holding decorated functions, and optional _install.py / _uninstall.py for setup and teardown. The loader recurses and auto-imports *.py, skipping anything starting with _ — so install scripts coexist with runtime code without colliding.

Install lifecycle:

aegis plugin install <name> --from gh:owner/repo#plugins/<name>
aegis plugin list / show / update / search / uninstall

--from accepts gh:owner/repo[@ref][#path] (HTTPS git archive fetch) or file:///abs/path (local copy). Without --from, the default registry is gh:apiad/aegis#plugins/. Installs roll back on failure; .aegis/plugins.lock records the resolved source + timestamp.

Two canonical plugins (in this repo)

skill-system — Claude-Code-style skill selection on any harness. A pre_turn hook injects a numbered menu parsed from .aegis/skills/*.md; a load_skill(name) @tool pulls the body on demand. ~100 lines of Python.

aegis plugin install skill-system --from gh:apiad/aegis#plugins/skill-system

memory-system — Hermes-inspired persistent memory with periodic dreaming. Per-project .aegis/memory/ holds a user-edited SOUL.md + USER.md, a MEMORY.md index over typed entries (user / feedback / fact / reference), and a dreams/ log. The pre_turn hook injects the persona bundle on turn 0 and top-K relevant teasers on later turns. Five memory_* @tools let the agent read/write/search. The dream @workflow runs a three-stage consolidate-plus-synthesize pass over the last week of session transcripts; install asks once whether to schedule it daily at 03:00. Exercises every substrate primitive (@hook, @tool, @workflow) end-to-end.

aegis plugin install memory-system --from gh:apiad/aegis#plugins/memory-system

Full protocol reference: Plugins documentation.

Remote plane — cross-machine handoff

aegis serve can expose a second HTTP plane — distinct from the loopback MCP plane, bound wherever you want it reachable — that other aegis serve instances POST into. One agent on one machine can hand a long task off to another without leaving the substrate:

aegis_enqueue(
    queue="implementation",
    payload="Implement the design at <path> with TDD…",
    from_handle="lucid-knuth",
    target="builder",           # ← new — routes to a remote aegis
)
# → {task_id: "01J…", target: "builder",
#    callback_note: "no wire return channel in v1; completion behavior
#                    is whatever the receiving serve is configured to do"}

Configuration lives in .aegis.yaml. Outbound — the peers this serve can call:

remotes:
  builder:
    url: http://100.64.0.5:8556
    # token: "<optional bearer>"

Inbound — opt-in receive side; default off:

remote_plane:
  bind: 100.64.0.5:8556         # the address to listen on
  accept_tokens: []             # optional bearer-token allowlist
  accept_from: []               # optional source-IP allowlist

Bind the plane wherever it should be reachable from, and only from — typically a private overlay network (Tailscale/Headscale/WireGuard/ VPN) so the network itself is the trust anchor. Bearer-token and source-IP gates compose with AND on top. By default the call is fire-and-forget: the receiving serve runs the worker under its own config and whatever it does on completion (commit and push, message through its own bridge, write to a shared folder, nothing) is up to it.

Since v0.8.0, aegis_enqueue(target=…, callback=True) opts in to a wire callback that delivers the remote worker's final message back to the originating agent's inbox as a normal ✉ from queue:<peer>:<name> envelope (symmetric peers config required — both sides define each other in remotes:). A small /remote/v1/schedule/* control plane (with aegis_schedule_* MCP tools and aegis schedule push --to <peer> / --remote <peer> CLI verbs) lets one serve push schedules into a peer and inspect or remove them remotely — useful for self-scheduling future work or managing a fleet from one host. Full surface, error model, and patterns in docs/remote.md.

Per-queue budgets

— the programmable pillar.

Declare rolling USD or output-token ceilings on any queue. The substrate enforces them at enqueue time — if admitting the task would push the queue over any configured ceiling, the enqueue is rejected immediately with a structured error that names the blocking constraint and gives an ETA for when the queue will unblock.

# .aegis.py
queues = {
    "impl": {
        "agent": "opus",
        "max_parallel": 2,
        "budgets": [
            {"usd": 1.00,             "window": "1h"},
            {"usd": 10.00,            "window": "24h"},
            {"output_tokens": 500000, "window": "1h"},   # runaway belt
            {"usd": 50.00,            "window": "7d"},
        ],
    },
    "fast": {
        "agent": "haiku-fast",
        "max_parallel": 4,
        # no budgets: key → no caps; behaves as before
    },
}

All-must-allow: a task is admitted only if every budget entry is under its ceiling. When rejected, the error names every blocking constraint, how much was spent vs the limit, and an unblock_at ETA.

Budget state is visible via aegis budget list/show (with --remote <peer> for cross-host), the aegis_budget_status MCP tool, and the read-only GET /remote/v1/budget endpoints on the remote plane. No alerts — observability is pull-only; the rejection at enqueue time is the only loud signal.

See docs/budget.md for the full model.

Scheduled workflows

— the programmable pillar.

Aegis runs a cron-style scheduler alongside QueueManager and the inbox router. Schedules are declared in .aegis.yaml and can be split into drop-in overlays under .aegis/schedules/<name>.yaml. Each entry names a workflow (built-in or registered), a trigger (cron or fire_at), a lifecycle (forever / once / {fires: N} / {until: <iso>}), and an overlap policy (skip / queue / kill).

# .aegis.yaml
schedules:
  morning-briefing:
    workflow: prompt
    cron: "0 6 * * *"
    timezone: America/Havana
    args: { agent: default, message: "Write today's briefing." }
  ci-watch:
    workflow: enqueue
    cron: "*/5 * * * *"
    lifecycle: forever
    on_overlap: skip
    args: { queue: ci, payload: "Check CI status and report failures." }

Two workflows ship in-tree: prompt (one-shot agent message) and enqueue (scheduler → queue handoff).

The substrate writes a JSONL audit log per schedule under .aegis/state/schedules/<name>.jsonl plus a derived schedules.snapshot.json for dashboards. On boot it replays each log to rebuild fire counts and closes any dangling fire_requested record as failed:interrupted. Editing .aegis.yaml or any overlay file hot-swaps the schedule table without a restart — entries that didn't change keep their state.

aegis schedule list                # current schedules + next fire
aegis schedule show morning-briefing
aegis schedule run morning-briefing   # force-fire once
aegis schedule disable morning-briefing  # comment-preserving YAML edit
aegis schedule logs morning-briefing -n 50

Install

pip install aegis-harness        # or: uv pip install aegis-harness

Requires Python 3.13+ and at least one of: claude, gemini, or opencode on your PATH, signed-in.

Quickstart

aegis          # full-screen TUI — first-class UI for local dev
aegis web      # installable PWA — first-class UI for remote (and local) dev

aegis has two co-equal first-class UIs over one backend: the full-screen TUI for local development, and an installable, mobile-first web/PWA client for remote development over a flaky link (and locally too). Both render the same transcripts with identical fidelity; sessions are shared across them. aegis web ensures a token, opens your browser, and serves the client.

With no .aegis.yaml in the directory, aegis drops you straight into the TUI ConfigPanel — press a to add your first agent and save. The scriptable equivalent:

aegis config agent add main --provider claude-code \
                            --model opus --effort high
aegis                       # start the TUI normally

.aegis.yaml is declarative YAML — edit it by hand or use aegis config verbs (every section reachable: agents, queues, default-agent, plugin-dir). Mid-session, reach the ConfigPanel via F2.

Keys

A backgrounded tab that finishes shows a * and rings the bell.

Configuration

.aegis.yaml is declarative YAML. Author it interactively (the TUI ConfigPanel — boot-into-panel when no config exists, F2 mid-session) or with the scriptable CLI (aegis config agent add, aegis config queue add, aegis config default-agent, …). Shape:

default_agent: default

agents:
  default:
    provider: claude-code
    model: opus
    effort: high
    permission: auto
  reviewer:
    provider: claude-code
    model: sonnet
    permission: read
  fast:
    provider: gemini
    model: gemini-3-flash-preview
    permission: full
  oss:
    provider: opencode
    model: opencode/kimi-k2.6
    permission: full

queues:
  review:
    agent: reviewer
    max_parallel: 2
  fast:
    agent: fast
    max_parallel: 4

Full reference: Configuration.

Headless + web

aegis serve runs the SessionManager and MCP plane without the TUI; add a web: block to serve the installable, mobile-first web/PWA client and drive the team from any browser:

# .aegis.yaml
web:
  bind: 127.0.0.1          # front with a reverse proxy for remote access
  port: 8899               # omit to auto-pick
  # token: "…"             # or set AEGIS_WEB_TOKEN (env wins) — keeps it out of git

aegis web ensures a token, opens your browser, and serves. The TUI, the web client, and (eventually) a remote TUI all speak the same WebSocket protocol over one backend, so sessions are shared across them. A systemd unit template lives at scripts/aegis-serve.service.

Troubleshooting: flicker / tearing in the TUI

If the cursor, the working spinner, and scrolling all flicker or stutter as if two frame rates are fighting — and it starts or stops when you plug in a monitor, close the laptop lid, or wake from sleep — this is not an aegis bug. aegis sets no custom render options, and Textual already negotiates the terminal's synchronized-output (DEC mode 2026) protocol so frames are emitted atomically. The flicker is one layer lower: your terminal emulator's GPU renderer pacing against a display whose refresh rate just changed. It affects every TUI (btop, htop, …), not just aegis, and the fix is a terminal/GPU setting.

Mitigations by terminal:

• GTK4/VTE terminals — Ptyxis, GNOME Console, Black Box, GNOME Terminal (GTK4). GTK removed its old GL renderer in 4.18; on older iGPUs (e.g. Intel HD 5xx / Skylake) the newer GPU renderers desync against the Wayland frame clock after a display change. Force the software renderer:

  mkdir -p ~/.config/environment.d
  echo 'GSK_RENDERER=cairo' > ~/.config/environment.d/50-gsk-renderer.conf
  # log out and back in — this must apply to the terminal's own GTK process,
  # not the shell inside it. Reversible: delete the file + re-login.
  

  Costs negligible CPU for a terminal. GSK_RENDERER=gl is a middle ground that keeps GPU accel. Refs: GNOME Discourse, GTK docs.

• kitty. Set sync_to_monitor no in kitty.conf — it stops tying redraws to the monitor refresh (the source of mixed-refresh tearing / CPU spikes). Ref: kitty #3154.

• WezTerm. Try front_end = "Software", or pin max_fps to your lowest monitor's refresh rate — WezTerm has known jitter/adaptive-sync issues on mismatched-refresh multi-monitor setups. Refs: max_fps, wezterm #2450, #7611.

• Alacritty. No built-in vsync toggle; use the Mesa env knob vblank_mode=0 alacritty (or 0/3 to taste) to change GL sync behaviour. Ref: alacritty #1739.

• Any terminal. Keeping the aegis window on a single monitor, or setting your displays to a matching refresh rate, avoids the mismatch that triggers it.

Docs

Full documentation: https://apiad.github.io/aegis/

• Install
• Usage
• Configuration
• Drivers — Claude / Gemini / OpenCode
• Queues — inter-agent delegation
• Canvas — shared markdown blackboard
• Terminals — shared live PTY
• Groups — broadcast-and-gather committees
• Remote plane — laptop ↔ VPS enqueue over HTTP
• Workflows — Python orchestration + catalog
• Budgets
• MCP plane — the tool surface
• Architecture
• Roadmap
• API reference

Status

Beta. Personal-infrastructure-grade, evolves fast. Expect change before 1.0. See the roadmap for what's next.

License

MIT — see LICENSE.